Meta introduces Muse Spark Open for local coding on Mac

Meta has launched Muse Spark Open, an open-weight model described by the company as the most capable in the United States. Designed to run locally on Mac, it targets complex coding and agents, featuring a context window of one million tokens and 816 billion parameters.

Demonstrated Uses in Coding and Video, but with Colossal Memory Requirements

Meta has integrated Muse Spark Open into agentic workflows, notably for a Mastodon client, where it was used to fix bugs, add features, and perform a complete build. The company also states that it has utilized the model to drive DaVinci Resolve, enabling the revision and cropping of video sequences. To run Muse Spark Open locally, at least 1 terabyte of RAM is required, with the optimal configuration being a Mac Studio equipped with the M5 Ultra chip.

Announced Capabilities: 816 Billion Parameters and 1 Million Tokens

Muse Spark Open incorporates 816 billion parameters and offers a context window that can reach one million tokens. This model supports multimodality from its design and targets advanced coding tasks as well as agentic work-related uses. It is intended to operate locally on Apple computers and has been associated with Muse Code, Meta's AI coding terminal.

Immediate Availability and Claimed Positioning in the United States

Muse Spark Open is now available through several API platforms. Meta presents it as the most capable open-weight model in the United States and claims, based on evaluations published on Thursday, that it surpasses other popular open-weight models. Open-weight models are touted as potentially more cost-effective, transparent, and flexible than closed models.